Research On Construction Technology Of 3D Semantic Map Of Mobile Robot

With the continuous development of AI technology,intelligent agents represented by unmanned vehicles,unmanned submersibles,and unmanned aerial vehicles have become research hotspots in academia and industry.An efficient semantic SLAM system is the basis for mobile robots to perform advanced tasks such as autonomous positioning,navigation,and human-computer interaction in an unknown environment.As a common core technology of the system,the construction of three-dimensional semantic map is a current difficult problem.In particular,the existing mature laser or visual SLAM systems are still unable to understand and apply high-dimensional semantic information in the environment.The implementation of the existing semantic SLAM system highly relies on powerful computing power and unexplainable deep learning theory,and is not suitable for vehicle-mounted embedded computing platforms for mobile robots.Therefore,on the basis of ORB-SLAM,this paper adopts the improved lightweight target detection module based on Yolov4 and the 3D semantic map construction method fused with 2D candidate frames to complete the mobile robot trajectory estimation while constructing a more accurate 3D semantic map.Its main work includes the following aspects.First,in view of the problem that the target detection algorithm of convolutional neural network requires high computing power of the computing platform,based on the single-stage detector Yolov4,a lightweight feature extractor network is redesigned,and the network structure and related hyperparameters are adjusted.And fusion meta-network scene classifier is used to improve the real-time performance of target detection;combine the public data set and the experimental environment data set to train the neural network to obtain the model weight file that meets the experimental requirements,and complete the semantic information extraction of the mobile robot environment based on this;Secondly,in view of the high computing power requirements based on the Yolov4 target detection algorithm,a meta-network scene classifier and a lightweight target detection network are designed to classify the detection scene and target detection in key frames.And for the current target recognition result that the candidate frame is axis-aligned,the angle and pose information of the object in the image cannot be effectively represented,and the problem of large errors caused by the semantic segmentation of the three-dimensional point cloud on this basis is proposed.A new end-to-end target detection method in any direction is proposed.This method is used to detect targets with a large aspect ratio in an image,and merge 2D target detection candidate frames to segment the 3D point cloud,and obtain a more accurate 3D point cloud segmentation and data-associated semantic map.Finally,this thesis builds a prototype system for the construction of a three-dimensional semantic map of a mobile robot to implement and verify the method proposed in this paper.Combined with the public data set and actual experimental data,the method proposed in this paper is verified and analyzed.The results show that the method proposed in this paper and the prototype system built on this basis are feasible and practical.